EV Sales Forecasting — Time Series Analysis

Forecasting annual electric vehicle sales across 40+ countries using IEA and World Bank data.

Overview

This project predicts country-level EV sales by modeling log growth rates with GradientBoosting, blended with a lag-1 baseline. It demonstrates a complete data science workflow: data integration, feature engineering, modeling, and evaluation.

Results

Split	Model	MAE	RMSE	MAPE	R²
Validation	Lag-1 baseline	71,696	343,459	37.0%	0.88
Validation	Growth GBR (blend=0.6)	29,456	127,047	23.3%	0.98
Test	Lag-1 baseline	80,350	445,102	31.7%	0.92
Test	Growth GBR (blend=0.6)	23,964	47,135	36.5%	0.999

Compared to the initial Ridge regression model (test MAPE 42.4%), the improved approach reduced MAE by 70% and RMSE by 86%.

Key Improvements

Change	Why it helped
Predict log growth rate instead of absolute sales	Makes the model scale-invariant — China (11M) and small markets (5K) contribute equally
GradientBoosting instead of Ridge	Captures non-linear feature interactions without manual engineering
Added macro features (GDP, inflation, unemployment)	Provides economic context beyond just historical sales
Added growth rate & market maturity features	Captures momentum and saturation effects
Tuned blend weight on validation set	Optimal weight (0.6) found via grid search, replacing hardcoded 0.5
Extended training data (≤2023)	More recent patterns improve generalization to 2024

Data Sources

Source	Description	Format
IEA Global EV Data Explorer	EV sales, stock, market share by country (2010–2024)	Excel
World Bank WDI	GDP per capita, CPI inflation, unemployment rate	CSV

See data/README.md for detailed data documentation.

Project Structure

├── data/
│   ├── raw/                    # Original datasets (see data/README.md)
│   └── README.md               # Data dictionary & download instructions
├── figures/                    # Generated charts (created by notebook)
├── notebooks/
│   ├── 01_ev_sales_forecasting_portfolio.ipynb   # Main analysis
│   └── ev_sales_forecasting_colab.ipynb          # Self-contained Colab version
├── src/
│   └── ev_sales_forecasting.py # Core pipeline
├── requirements.txt
├── .gitignore
├── LICENSE
└── README.md

Quick Start

Option A — Run locally:

git clone https://github.com/Yemyu/ev-sales-forecasting.git
cd ev-sales-forecasting
pip install -r requirements.txt

# Download data (see data/README.md) and place in data/raw/
jupyter notebook notebooks/01_ev_sales_forecasting_portfolio.ipynb

Option B — Run on Google Colab:

Upload notebooks/ev_sales_forecasting_colab.ipynb to Google Colab, then follow the in-notebook prompts to upload the 5 data files. No local setup needed.

Methodology

1. Data integration — Merge IEA EV panel with World Bank macro indicators via fuzzy country-name matching
2. Feature engineering — Lag-1/2 sales, lag-1 stock (log-transformed), year trend, country fixed effects, GDP, inflation, unemployment, sales growth rate, share change, stock-sales ratio
3. Modeling — GradientBoosting on log growth rates, blended with lag-1 baseline (weight tuned on validation set)
4. Evaluation — Train ≤2021 → validate 2022–2023 (tune blend weight) → retrain ≤2023 → test 2024

Limitations & Future Work

• Test MAPE (36.5%) still above naive baseline (31.7%) — small markets with rapid structural shifts drive percentage errors
• Could explore LightGBM / XGBoost with Bayesian hyperparameter tuning
• Policy variables (subsidies, emission standards) would help capture regulatory shocks
• Per-region or per-market-tier models may improve heterogeneous markets

Tech Stack

Python · pandas · scikit-learn · matplotlib · seaborn

License

MIT